Hindawi

International Journal of Aerospace Engineering

Volume 2018, Article ID 9570873, 12 pages
https://doi.org/10.1155/2018/9570873

Research Article
Aircraft Gas Turbine Engine Health Monitoring System by Real
Flight Data

Mustagime Tülin Yildirim1 and Bülent Kurt 2

1
Department of Aircraft Electrical and Electronics, Faculty of Aeronautics and Astronautics, Erciyes University,
38039 Kayseri, Turkey
2
Aircraft Technology Program, Erzincan University, 24100 Erzincan, Turkey

Correspondence should be addressed to Bülent Kurt; bulentkrt@gmail.com

Received 3 July 2017; Revised 7 November 2017; Accepted 10 December 2017; Published 14 March 2018

Academic Editor: Angel Velazquez

Copyright © 2018 Mustagime Tülin Yildirim and Bülent Kurt. This is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.

Modern condition monitoring-based methods are used to reduce maintenance costs, increase aircraft safety, and reduce fuel
consumption. In the literature, parameters such as engine fan speeds, vibration, oil pressure, oil temperature, exhaust gas
temperature (EGT), and fuel flow are used to determine performance deterioration in gas turbine engines. In this study, a new
model was developed to get information about the gas turbine engine’s condition. For this model, multiple regression analysis
was carried out to determine the effect of the flight parameters on the EGT parameter and the artificial neural network (ANN)
method was used in the identification of EGT parameter. At the end of the study, a network that predicts the EGT parameter
with the smallest margin of error has been developed. An interface for instant monitoring of the status of the aircraft engine has
been designed in MATLAB Simulink. Any performance degradation that may occur in the aircraft’s gas turbine engine can be
easily detected graphically or by the engine performance deterioration value. Also, it has been indicated that it could be a new
indicator that informs the pilots in the event of a fault in the sensor of the EGT parameter that they monitor while flying.

1. Introduction
Aircraft engines produce the power needed for aircraft. For
this reason, aircraft engines are very important for flight
safety [1]. Today’s complex and advanced technology sys-
tems require advanced and expensive maintenance strategies
[2]. Maintenance services are costly for airline companies.
For manufacturers, maintenance is a source of revenue.
According to Dennis and Kambil, though after-sales service
and parts sales constitute 25% of the manufacturer’s income,
it makes up 40–50% of company profits [3].
Because of the high cost of maintenance, gas turbine
engines must be operated within specified physical limits [4].
Today’s aircraft engines are made safer by increasing the num-
ber of control parameters and sensors [5]. The engines have a
complex mechanical system. Because aircraft engines operate
at high temperatures, high pressures, and high speeds, there
are lots of possibilities of various faults in the aircrafts [6].
Gas turbine engines show the effects of wear and tear over
time. A small fault during the flight does not prevent the
engine from running, but if this fault is not detected, it could
lead to a bigger fault. If these bigger faults cannot be prevented
in the aircraft, it can lead to high maintenance costs and
accidents. When aircrafts are taken for maintenance, the con-
dition of the gas turbine engine is investigated by various tests
and measurements [7]. Condition-based maintenance (CBM)
is being performed to provide effective and efficient mainte-
nance in today’s maintenance services. In the literature, there
are major developments in the CBM method with studies on
machine condition monitoring and fault diagnostics [8]. A
typical CBM program consists of three steps [9]:
(i) Data acquisition step
(ii) Data processing step
(iii) Maintenance decision-making step
2 International Journal of Aerospace Engineering
In a CBM program, we can group two important parts as
diagnostics and prognostics. With diagnostic monitoring,
malfunctions or abnormal conditions are detected when a
system or its subsystems are in operation [10]. With the
prognostics approach, the current state of the motor and
possible future failures can be predicted. The following diag-
nostic systems have been applied on an aircraft: the typical
diagnostic procedure consists of three steps: observation,
comparison, and diagnosis, which are shown in Figure 1 [11].
In the literature, advanced engine condition monitoring
methods can be classified as a model-based method (such
as equality equation, parameter prediction, gas path analysis,
nonlinear least square, and Kalman filter) and soft comput-
ing method (such as artificial neural networks, fuzzy logic,
rough sets, decision tree, and expert systems).
In the first phase of the study, some sample studies on
failure prediction in gas turbine engines have been men-
tioned. In the second phase, the main operating parameters
of the aircraft gas turbine engines and the importance of
the EGT parameter are discussed. In the third phase, multiple
regression analysis is performed between the EGT parameter
and the aircraft performance parameters using the SPSS pro-
gram. In the next phases, N1 speed, N2 speed, pitch, angle of
attack, roll, vertical acceleration, total air temperature, and
ground/air variables from aircraft performance parameters
are taken as input parameters to the artificial neural net-
works, and prediction of EGT parameters has been done in
MATLAB Simulink environment to monitor the condition
of an aircraft gas turbine engine.
2. Literature
There are many studies in the literature where different
methods are used to monitor the condition of gas turbine
engines used on airplanes. Zhang et al. predicted failure of
aircraft engines using nonlinear adaptive estimators [12],
Pourbabaee et al. predicted sensor failures in gas turbine
engines using multiple model-based methods [13], Zhao
et al. made a prediction of failure in the gas turbine engines
using fuzzy matrix and the principle of maximum member-
ship degree [14], Afghan et al. have designed an expert sys-
tem to detect faults for gas turbine combustion chambers
[15], and Zaidan et al. have made failure prediction in gas
turbine engines with Bayesian hierarchical models [8]. Lu
and his colleagues performed a turbofan engine status mon-
itoring with extended Kalman filter [16], Amozegar et al.
have made failure prediction in gas turbine engines using
dynamic neural network [17], Zhou et al. used the support
vector machine method to diagnose gas turbine engine fail-
ures [18], and Chacartegui et al. have worked on real-time
simulations of medium-sized gas turbines [19].
3. Main Operation Parameters of the Engine
There are many electronic indicators and systems that help
pilots on board during flight. The pilots monitor the indica-
tors, have knowledge of the current status of the aircraft sys-
tems, and carry out the flight. When the values appearing on
the indicators indicate that they are out of the normal
Observation
Comparision
of observed
Plant Diagnosis
and expected
behavior
Figure 1: Typical diagnostic procedure [27].
working range or that there is a malfunction in the systems,
the pilots are able to request help from the operation services
via the communication systems, or in cases where they are
not considered to be very serious, pilots record them in the
flight book and let them be added to the maintenance pro-
gram. Pilots must monitor the main operating parameters
while controlling the aircraft and making decisions. In
Figure 2, fan speed (N1 speed) and exhaust gas temperature
(EGT), the gas turbine engine’s main operating parameters,
are shown. The fan speed is usually used as a push indicator,
while the EGT is the parameter that allows the gas turbine
engine health to be monitored. In some engine models,
engine pressure ratio (EPR) and N2/N3 speed are used to
monitor thrust [5]. The importance of EGT parameter is
given below.
3.1. Exhaust Gas Temperature (EGT) and EGT Margin. The
EGT is the temperature of the gas turbine engine’s exhaust
in degrees and indicates the performance status of the gas
turbine engine’s design. High EGT values result in further
deterioration and wear in the engines. Table 1 shows which
failures and causes will lead to high EGT values in engines.
In addition, pilots should pay attention to the EGT value,
especially at takeoff, because exceeding the EGT limit may
cause damage to the gas turbine engine components or
reduce the service life of components. It is therefore impor-
tant to keep the EGT as low as possible [20]. At the maxi-
mum EGT temperature, the gas turbine engines can reach
at takeoff and has a certificate showing a red reference line.
An example is shown in Figure 3. The difference between
the highest EGT temperature reached during the takeoff and
the EGT determined by the red reference line is called EGT
margin (Figure 3) [21]. EGT margin is used to obtain infor-
mation about motor performance.
EGT margin deterioration is due largely to engine com-
ponents. Compressor fouling, seal leakage, increased tip
clearances, and airfoil erosion are among the deterioration
reasons. EGT margin deterioration on airplanes causes the
increase in engine thrust, the decrease in engine usage, the
decrease in average flight distance, and the difficulty of the
working environment. Figure 4 shows some sample faults
that may increase the EGT value.
The EGT margin value is a parameter that is controlled
by experts only when in maintenance at the takeoff and is
used to evaluate the condition of the gas turbine engine.
The more the difference between the max EGT value
International Journal of Aerospace Engineering 3

N1 speed (exhaust
gas
temperature)
N2/N3 speed

EGT

EPR

Figure 2: Main operation parameters of aircraft engines [20].

Table 1: Gas turbine engine failures and causes [28].

Fuel Reverser
Engine Severe Bird ingestion/ Tailpipe Hot Fuel
Surge Seizure Flameout control Fire Icing inadvertent
separation damage FOD fires start leak
problems deploy
High
X X O O X O X O
EGT
X: symptom very likely; O: symptom possible. Blank fields mean that the symptom is unlikely.

EGT redline

EGT margin °C = EGT redline − EGT takeoff

To = 850 °C
EGT °C

EGT margin = 950 °C − 850 °C = 100 °C RL = 950 °C

EGT
850 °C

Full takeoff power

Time since initiation of takeoff

Figure 3: EGT margin.

reached at takeoff and the EGT redline determined by the

manufacturer is, the better the performance of the gas turbine
engine. This assessment is valid only for takeoff not for other
flight stages.

4. Method
In this study, by comparing the EGT value that should be at
the takeoff, cruise, and landing phase of the aircraft and the
EGT value obtained from the sensor, RMS value is obtained,
and this value displays the deterioration in performance. For
this purpose, multiple regression analysis (MRA) and artifi-
cial neural networks (ANN) have been chosen to predict
the EGT values obtained from real flight data. MATLAB pro-
gram for artificial intelligence analysis and SPSS Statistics 22
package program for statistical regression analysis were used.
In MRA and modeling with ANN, 12,935 samples were used.
70% for training, 15% for validation, and 15% for testing
were used from these samples. The data used for modeling
was recorded during the flight of the aircraft, and no changes
were made to the data. Training, validation, and test data
were randomly selected from the dataset. The same training,
validation, and test data were used in both analyses to evalu-
ate the comparison and modeling. The training, test, and val-
idation data statistical parameters used in the prediction of
EGT parameters are given in Tables 2–4, respectively.
Figure 4: Sample faults in aircraft engines.
4.1. Statistical Regression Analysis. In the regression, the goal
is to find relationships between the descriptive (independent)
variables and the described (dependent) variable. As y is a
dependent variable, x1, x2,…, xk are independent variables,
the method which expresses the causal relationship between
variables as a mathematical model is called a multiple linear
regression model [22]. In multivariable linear regression
analysis, the general formula of the regression equation that
gives the shape of the relationship between variables can be
expressed by
y = β0 + β1 x1 + β2 x2 + ⋯ + βk xk + ε, 1
where y, x1 , x2 , … , xk represent the observable values and βj,
j = 0, 1, … , k, as the regression coefficients βk parameters as
4 International Journal of Aerospace Engineering

Table 2: Training data range used in the prediction of EGT.

Statistics
Angle of Vertical Total air Exhaust gas
N1 speed N2 speed Pitch Roll Ground/air
attack acceleration temperature temperature
N
Valid 1005 1005 1005 1005 1005 1005 1005 1005 1005
Missing 0 0 0 0 0 0 0 0 0
Mean 60.9772139 81.6708458 2.112438 0.726965 −0.883582 0.995590 −3.034229 0.87 584.232836
Median 66.9000000 84.4000000 1.100000 −0.400000 −1.100000 0.994000 −3.300000 1.00 574.500000
Std. deviation 25.43270881 10.67348876 3.9393376 1.7996899 4.1258416 0.0465146 17.7587592 0.341 116.2669986
Range 75.60000 38.30000 22.9000 9.5000 51.7000 0.5470 84.8000 1 433.5000
Minimum 20.40000 59.00000 −3.2000 −1.8000 −21.8000 0.7750 −24.5000 0 432.5000
Maximum 96.00000 97.30000 19.7000 7.7000 29.9000 1.3220 60.3000 1 866.0000

Table 3: Test data range used in the prediction of EGT.

Statistics
Angle of Vertical Total air Exhaust gas
N1 speed N2 speed Pitch Roll Ground/air
attack acceleration temperature temperature
N
Valid 215 215 215 215 215 215 215 215 215
Missing 0 0 0 0 0 0 0 0 0
Mean 59.8427907 81.3158140 1.969767 0.838140 −1.316744 0.996465 −3.040000 0.90 576.039535
Median 61.9000000 83.1000000 1.100000 .000000 −1.100000 0.997000 −2.000000 1.00 572.000000
Std. deviation 25.30062671 10.45487550 3.9448260 1.9207515 3.5105979 0.0413821 17.1693103 0.304 115.2838515
Range 75.60000 37.80000 21.8000 9.9000 50.3000 0.3550 53.0000 1 427.5000
Minimum 20.40000 59.30000 −2.8000 −1.1000 −21.8000 0.8710 −24.5000 0 434.0000
Maximum 96.00000 97.10000 19.0000 8.8000 28.5000 1.2260 28.5000 1 861.5000

Table 4: Validation data range used in the prediction of EGT.

Statistics
Angle of Vertical Total air Exhaust gas
N1 speed N2 speed Pitch Roll Ground/air
attack acceleration temperature temperature
N
Valid 215 215 215 215 215 215 215 215 215
Missing 0 0 0 0 0 0 0 0 0
Mean 59.6409302 80.8041860 1.791628 0.672093 −1.298140 1.000530 −4.323721 0.86 575.411628
Median 65.9000000 83.9000000 1.100000 −.400000 −1.400000 0.997000 −8.300000 1.00 572.000000
Std. deviation 25.67176527 10.77929528 3.6772253 1.7983021 4.4961766 0.0397756 17.7368312 0.347 108.3661869
Range 75.30000 37.80000 22.5000 7.1000 51.0000 0.4170 64.0000 1 421.5000
Minimum 20.60000 59.30000 −2.8000 −1.1000 −21.8000 0.8360 −24.0000 0 442.0000
Maximum 95.90000 97.10000 19.7000 6.0000 29.2000 1.2530 40.0000 1 863.5000

the expected change in y for the unit changes in xk’s and e
error. The model investigated for regression analysis in this
study is given in Table 5.
Where xi values are, respectively, N1 speed, N2 speed,
pitch, angle of attack, roll, vertical acceleration, total air tem-
perature, ground/air, and β0 , β1 , β2 , β3 , β4 , β5 , β6 , β7 , β8 are
unknown parameters. The investigation of whether the coef-
ficients of the regression model are significantly different is
carried out by the t-test. When the significance value (Sig.)
of the t-statistic is less than 0.05, the assumption is accepted
and it is concluded that the coefficients in the model are
meaningful. Sig. value is the value used with the purpose of
determining the existence of statistical significance and the
level of proof of the difference, if any. The level of significance
is usually taken as 0.05 and 0.01. A level of 5% significance
means a confidence interval of 95%, and a level of
International Journal of Aerospace Engineering
Table 5: Investigational model for the regression analysis.
Dependent variable (y) Independent variable (xi)
Exhaust gas temperature (EGT) = β0 + β1 x1 + β2 x2 + β3 x3 + β4 x4
+ β5 x5 + β6 x6 + β7 x7 + β8 x8
Table 6: The interpretation of the Sig. value [23].
Sig. value Meaning
0.01 < Sig. ≤ 0.05 The results are significant.
0.001 ≤ Sig. < 0.01 The results are highly significant.
Sig. < 0.001 The results are very highly significant.
The results are considered not
0.05 ≤ Sig. < 0.10
statistically significant.
A trend toward statistical
Sig. > 0.10
significance is sometimes noted.
significance of 1% means a confidence interval of 99%. The
general approach to the interpretation of the Sig. value is
given in Table 6 [23].
4.2. Artificial Neural Networks. ANN are nonlinear modali-
ties that are composed of artificial neurons connected with
each other and that contain an input set and a single output.
In ANN, there are many cells and many bonds between
inputs and outputs. ANN can be divided into several subclus-
ters. These small groups are called layers. The network is
formed by hierarchically connecting layers. In ANN, learning
takes place in two stages. In the first step, the random weight
values are taken and the net output is determined for the
sample shown to the network. The weights are regenerated
by feedback or forward feed according to the response suit-
ability. In the second step, the weight values are changed by
showing different samples to the network in order to find
the best weight value that the correct output can obtain.
There are different algorithms used in the learning process.
The backpropagation algorithm is the most common. In
the backpropagation algorithm, it tried to reduce backward
errors. The training algorithm specifies how to adapt the
learning rule to ANN according to the nature of the problem
at hand. The samples selected during training should be
selected from each region of the problem space to represent
the problem space. The input and corresponding output vec-
tor are used to train the network. The output obtained from
each iteration result is compared to the target, and network
training is continued with the weight renewal process or the
training is terminated depending on the given error. After
the weights, providing the best output data is determined
in artificial neural networks, the processes carried out to
evaluate the learning status of the network are defined as
testing the network. In this phase, the output is produced
by using the most suitable weight values determined after
the training by introducing not seen samples to the network.
This output reveals the learning success of the network. The
more successful the result, the better the training perfor-
mance of the network.
5
4.3. Measurement Accuracy of Prediction Models. The accu-
racy of a prediction model depends on how close the predic-
tion values are to the actual observation value. If the model is
successful in predicting actual values, the prediction error
will be relatively low. The methods used in this study were
analyzed by three different error performance methods. The
first of these is the R2 determination coefficient and can be
expressed by
n
〠i=1 yi − xi 2
R2 = 1 − n
, 2
〠i=1 yi − y 2
where n is the number of observations, y is the actual values,
x is the predicted value, and y is the average actual value. The
other error analysis method is expressed by the mean abso-
lute percentage error (MAPE).
n
〠i=1 et /yt
MAPE = 100 % 3
n
In this formula, et = yt − xt , yt is the actual observation
value, xt the prediction value, n is number of observations
in the prediction period, and et is the prediction error in t
period. Another error analysis method is expressed by the
mean square error (MSE).
1 n
MSE = 〠 y − xi 2 , 4
n i=1 i
where xi is the predicted value of the ith unit, and yi is the
actual value of the ith unit. The closer the predicted values
are to the actual values, the smaller the MSE becomes; the
further away from the actual values the greater the MSE
is. Root mean square error (RMSE) in (5) is a frequently
used measure of the differences between values (sample
values) predicted by a model or an estimator and the values
actually observed.
1 n
RMSE = 〠 y − xi 2
5
n i=1 i
5. Results and Discussions
5.1. Determination of Effective Aircraft Performance
Parameters and Prediction by Multiple Regression Analysis.
Multiple regression analysis was performed to determine
the effect of independent variables on dependent variable
EGT parameter. When the significance value (Sig.) of t-statis-
tic is greater than 0.05 in multiple regression models, the
analysis is repeated by subtracting the meaningless variable
from the model. A new model has been derived by subtract-
ing meaningless variables from the data recorded during
flight of the aircraft. In the obtained model, N1 speed, N2
speed, pitch, angle of attack, roll, vertical acceleration, total
air temperature, and ground/air parameters were found to
be important parameters affecting EGT parameter
(Table 7). Here, EGT is a dependent variable, N1 speed, N2
speed, pitch, angle of attack, roll, vertical acceleration, total
air temperature, and ground/air are independent variables,
6 International Journal of Aerospace Engineering

Table 7: Multiple regression analysis results for EGT training data.

Coefficientsa
Unstandardized
Model Standardized coefficients Correlations
coefficients t Sig.
B Std. error Beta Zero order Partial Part
1
(Constant) 854.361 32.679 26.144 0.000
N1 speed 7.279 0.261 1.592 27.924 0.000 0.718 0.663 0.217
N2 speed −7.605 0.529 −.698 −14.381 0.000 0.653 −0.415 −0.112
Pitch 8.654 0.553 0.293 15.651 0.000 0.874 0.444 0.122
Angle of attack 2.139 0.799 0.033 2.675 0.008 −0.003 0.084 0.021
Roll 1.592 0.238 0.056 6.687 0.000 −0.139 0.207 0.052
Vertical acceleration −72.026 20.637 −0.029 −3.490 0.001 −0.107 −0.110 −0.027
Total air temperature 3.418 0.146 0.522 23.377 0.000 0.003 0.595 0.182
Ground/air −33.776 5.793 −0.099 −5.831 0.000 0.007 −0.182 −0.045
a
Dependent variable: Exhaust gas temperature.

Table 8: Model summary.

Model summaryb
Change statistics
Model R R square Adjusted R square Std. error of the estimate
R square change F change df1 df2 Sig. F change
1 0.969a 0.940 0.939 28.6701942 0.940 1939.432 8 996 0.000
a b
Predictors: (constant), ground/air, vertical acceleration, pitch, roll, angle of attack, total air temperature, N2 speed, and N1 speed. Dependent variable: exhaust
gas temperature.

Table 9: The interpretation of The Pearson correlation coefficient attack, roll, vertical acceleration, total air temperature, and
[29]. ground/air parameters are highly correlated with dependent
variable EGT parameter (R = 0.969). Independent variables
Correlation value Meaning explain 93.9% (adjusted R square = 0.939) of the total vari-
0.00–0.19 Very weak ance (change) in the dependent variable EGT (see Table 8).
0.20–0.39 Weak According to the standardized regression coefficient (beta),
0.40–0.59 Moderate the relative importance of the independent variables over
0.60–0.79 Strong the EGT parameter is N1 speed, total air temperature, pitch,
roll, angle of attack, vertical acceleration, ground/air, and N2
0.80–1.0 Very strong
speed. When the t-test results on the significance of regres-
sion coefficients are examined, it is seen that N1 speed
B is unregulated regression coefficient, beta is standardized parameter is the most important (significant) independent
regression coefficient, Sig. is significance level, R is correla- variable on EGT. According to the results of regression
tion coefficient, and R2 is the determination coefficient. analysis, multiple regression equation’s mathematical model
According to Table 8, N1 speed, N2 speed, pitch, angle of for EGT prediction is expressed by

EGT = 854, 361 + 7, 279 N1 speed − 7, 605 N2 speed + 8, 654 Pitch + 2, 139 AOA + 1, 592 Roll
6
− 72, 026 Vertical acceleration + 3, 418 TAT − 33, 776 Ground/air

The analysis of the linear multiple regression model is
shown in Table 7. The fact that Sig. values of the parameters
used in this model are less than 0.01 means that it is a highly
significant model according to Table 6.
As a result of the correlation analysis, whether there is a
linear relationship and the degree of this relation if there is
any is calculated by a correlation coefficient, the correlation
coefficient takes values between −1 and +1. If there is a
positive correlation, the value of the variable y increases
as the value of the variable x increases, or the value of
the variable y tends to decrease as the value of the variable
x decreases. Negative correlation (negative relationship)
International Journal of Aerospace Engineering
Table 10: ANOVA results for EGT training.
Model Sum of squares
1
Regression 12753394.90
Residual 818692.113
Total 13572087.02
a
Dependent variable: exhaust gas temperature. bPredictors: (constant), ground/air, vertical acceleration, pitch, roll, angle of attack, total air temperature, N2
speed, and N1 speed.
Table 11: MRA error performance values.
MRA R2 RMSE MAPE % MAPE
Train 0.939 28.67 0.039463 3.9463
Test 0.947038 26.468 0.038119 3.8119
Validation 0.93702 27.131 0.036104 3.6104
means that if one of the variables increases, the value of the
other variable decreases.
When the values in Table 7 are interpreted according
to Table 9, it could be seen that there is a strong, positive,
and significant relationship between the pitch parameter
and the EGT parameter and that there is a strong positive
correlation between the N1 speed and N2 speed parameters
and the EGT parameter. It could also be seen that there is a
very weak, positive, and significant relationship between total
air temperature and ground/air parameters and EGT
parameters and that the angle of attack, roll, and vertical
acceleration seem to be very weak and negative and have a
significant relationship.
The fact that the Sig. value shown in Table 10
(Sig. = 0.000) is smaller than 0.001 indicates that the devel-
oped model is very highly significant. According to 6, EGT
parameter, training, test, and validation data are predicted
by multiple regression analysis. The error performance
values between the measured values and the predicted values
for the training, test, and validation stages are determined
according to 2, 3, 4, and 5. The comparison of EGT values
predicted by multiple regression analysis and EGT value
from the aircraft sensor at training, test, validation stage is
shown in Table 11.
5.2. Prediction with Artificial Neural Network. In the ANN
model; N1 speed, N2 speed, pitch, angle of attack, Roll,
vertical acceleration, total air temperature, and ground/air
are given as the input data to the network and EGT is
predicted as the output data. All the data have been pre-
sented to the network as it was received from the aircraft
without any change.
The limit values of the input and output parameters of
the ANN model are shown in Tables 2–4. 12,935 data
obtained from the flight of the aircraft were used in the
MATLAB program with “dividerand” command 70% of
which for training, 15% of which for testing, and 15% of
which for validation. The feed forward network type was
chosen as the artificial neural network model to be used in
7
ANOVAa
df Mean square F Sig.
8 1594174.363 1939.432 0.000b
996 821.980
1004
practice. In the feed forward calculation, the most important
event is trying to predict the output values with the lowest
MSE error margin. The model shown in Figure 5 is obtained by
changing the number of iterations and the number of neu-
821.98
rons in the hidden layer for the optimal network architecture
700.6043 for predicting the EGT parameter with the experiments per-
736.1433 formed. The best results of 11 different training functions to
predict the EGT parameter with the smallest error are shown
in Table 12. For experiments, a computer with Intel (R) Core
(TM) i5-4590 CPU @ 3.30 GHz 4 GB RAM was used.
When the results in Table 12 are compared against the
test MSE value, it is determined that the Levenberg-
Marquardt backpropagation function (trainlm) is the best
resultant. In Figures 6–8, it illustrates neural network regres-
sion plots for training, validation, and test. The correlation
coefficient of the EGT value from the aircraft sensor and the
predicted values of training, test, and validation phases are
higher (R > 0.99). In order to evaluate the performance of
the developed network, the error performance values between
measured values and predicted values for training, validation,
and test phases were determined according to 2, 3, 4, and 5
and the results are presented in Table 13. In addition, the net-
work characteristics developed are given in Table 14.
When the results are evaluated, determination coefficient
(R2 > 0.99) could explain more than 99% of EGT parameter
at training, testing, and the validation stage with N1 speed,
N2 speed, pitch, angle of attack, roll, vertical acceleration,
total air temperature, ground/air parameters. The closer the
R2 parameter is to 1, the better the performance of the
network. Lewis has classified the models with a MAPE value
of less than 10% as “very good,” models between 10% and
20% as “good,” models between 20% and 50% as “accept-
able,” and models higher than 50% as “wrong and faulty”
[24, 25]. When we look at the MAPE values of the
enhanced network, it is suggested that the network is
predicated on high accuracy since MAPE is <1% during
the training, testing, and validation phases. It also indi-
cates that the smaller the RMSE parameter, the higher
the performance of the network.
5.3. Modeling of the Engine Health Monitoring Program in
MATLAB Simulink Environment. In order for our system
to work instantly, we need to design an interface in the
MATLAB Simulink environment with the best ANN model
we have obtained as a result of the experiments. Figure 9
shows the structure of the ANN model modeled in MATLAB
Simulink environment.
8 International Journal of Aerospace Engineering

Input layer
N1 speed Hidden
N2 speed layers
Output
Pitch layer
AOA Predicted
EGT
Roll
Vertical acceleration
TAT Neurons
Ground/air Weights

Figure 5: Architecture of the proposed neural model.

Table 12: Comparison of the feedforward neural network training algorithms for their test MSE value.

Training algorithm Elapsed time (sec) Training R Validation R Test R Training MSE Test MSE
trainlm 4.203957 0.99767 0.99797 0.99719 62.8501 74.6465
traincgb 4.565835 0.99586 0.99299 0.99656 111.4932 91.2834
trains 10.047608 0.99561 0.99517 0.99653 118.4071 91.8717
trainrp 57.118179 0.99458 0.99297 0.9963 146.1263 97.9763
traincgp 3.699402 0.99401 0.99515 0.99501 161.8084 133.9647
trainscg 3.458904 0.99381 0.99084 0.99474 166.9843 139.1074
trainbfg 13.748363 0.99213 0.99176 0.99394 211.9474 160.1871
traincgf 7.667996 0.99179 0.99176 0.99395 222.3323 164.4613
trainoss 4.432661 0.99053 0.99197 0.99183 255.8224 215.3883
traingdx 3.614869 0.98838 0.98893 0.9893 312.1862 285.3237
traingda 4.008302 0.98312 0.98418 0.98733 459.2168 342.6505

Training: R = 0.99767 Validation: R = 0.99797

850 850

800 800

750 750
Output ≌ 1⁎Target + 0.7
Output ≌ 1⁎Target + 2.6

700 700

650 650

600 600

550 550

500 500

450 450
450 500 550 600 650 700 750 800 850 450 500 550 600 650 700 750 800 850
Target Target
Data Data
Fit Fit
Y=T Y=T

Figure 6: Neural network regression plots for training. Figure 7: Neural network regression plots for validation.
International Journal of Aerospace Engineering 9

Test: R = 0.99719 Differential analysis model needs to be modeled in

850 MATLAB Simulink environment. The interface designed
for the aircraft’s engine health monitoring in the MATLAB
800 Simulink environment is shown in Figure 10.
When the graph shown in Figure 11 is examined, the
750 values shown in the pink color represent the data obtained
from the sensor of the aircraft, the values shown in the yellow
Output ≌ 1⁎Target + 3

700 color represent the EGT values predicted by the ANN, and
the values shown in green indicate the motor performance.
650 The limit determined for the alarm level is shown in red.
The alarm level assumed for this aircraft engine is EGT
600 margin value. In order to make a system monitoring, values
obtained from the system needs to be used. Two types of data
550 are used in the applications:

500 (1) Data gathered from normal and faulty operation

conditions.
450
(2) Data gathered only from monitoring of normal con-
450 500 550 600 650 700 750 800 850 ditions. Collecting faulty data from gas turbine
Target engines like systems are very difficult and expensive.
Data
Due to aviation rules like EASA Part-M, aircrafts
Fit with only a completed maintenance and with a
Y=T Certificate of Release to Service are allowed to fly.
An aircraft with a completed maintenance means a
Figure 8: Neural network regression plots for test. collection of systems with a healthy operation. There-
fore, condition monitoring systems are constructed
Table 13: ANN error performance values (fsor trainlm). only for normal flight data.
ANN R2 MAPE % MAPE RMSE Since we used data from a real aircraft, only normal oper-
Train 0.99534601 0.00917102 0.917102 7.927808 ation data is in hand. It is apparent from simulation results
Test 0.99435716 0.002224188 0.2224188 8.639823 that instantaneous performance of the gas turbine engine
Validation 0.99594228 0.001924502 0.192450201 6.886876 can be monitored from takeoff to landing. In this study, we
compare EGT values one from instant prediction we did
and another gathered from the EGT sensor of the aircraft.
Even though there exists a difference in between, this is
Babbar et al. evaluated the difference between EGT tem- evaluated as an abnormal situation.
perature and flight parameters of two symmetric engines at In Figure 12, a fault detection example is shown. Gas
various stages of the flight to examine the condition of the turbine engines are composed of various components. A fail-
aircraft [26]. By differential analysis, it is aimed to compare ure on a single or multiple components will result in a perfor-
EGT value obtained from ANN with real EGT value coming mance defection in those engines. Corrosion, oxidation,
from the sensor of the aircraft and to easily notice when there damaged bearings, high blade clearance, damaged or burned
is any problem. In this study, it is necessary to instantly com- turbine blade and vanes, damaged combustors, damaged fuel
pare the values in order to evaluate the condition of the gas nozzles, and broken rotor discs and blades are some of
turbine engine. In this analysis, EGT.1 was determined as the failures that lead to performance degradation. Perfor-
the actual EGT value from the sensor of aircraft and EGT.2 mance degradation will result a change in EGT value. This
was determined as the parameter predicted by ANN. The study is not supposed to determine where and why a failure
similarity scale used to determine the status of the gas turbine has occurred. Exceeding EGT margin (+100°C) is consid-
engine is illustrated in ered as an abnormal situation. In this study, it is shown
that the performance of the aircraft can be monitored from
EGTdif f i = EGT 1 i − EGT 2 i 2
7 takeoff to landing.
Pilots follow various parameters during operation of the
Using EGT differences of each value, 8 is used to calculate aircraft. EGT parameter is one controlled during the takeoff.
the mean square root of the given flight phase. Engine manufacturers determine maximum EGT thresholds
during a takeoff. If the pilot does not track EGT, there is no a
1 n secondary system that generates warnings. If the engine
RMS Flightmode = 〠 EGTdif f i 8
n i=1 exceeds the maximum allowed EGT values, it will experience
serious failures with high probability. The system developed
The RMS value is shown as THE aircraft’s performance in this study will warn the pilot in case he misses the EGT
deterioration value in the interface program. value from the engine and warning display system.
10 International Journal of Aerospace Engineering

Table 14: Details of the neural model.

Network parameters Details

Architecture 8 inputs, 1 outputs, and 4 hidden layers with 11, 8, 4, and 3 neurons in each layer (11-8-4-3)
Training subset: 70% randomly selected recorded data (9045 patterns)
Set Validation subset: 15% randomly selected recorded data (1935 patterns)
Test subset: 15% randomly selected recorded data (1935 patterns)
Activation For hidden layers: tangent sigmoid tangent
Function For output layer: linear
Training algorithm Levenberg-Marquardt
Performance function criteria Minimum MSE
Stopping Validation stop
Criteria (Training is stopped when the validation error starts increasing)

Input Process input 1 Layer 1 a{1}

Layer 2 a{2}

a{1} Layer 3 a{3}

a{2} Layer 4 a{4}

a{3} Layer 5

a{4} Process output 1 Output

Figure 9: Modelling of the developed ANN model in MATLAB Simulink environment.

Input NNET Output

My data

Predicted EGT
From
workspace
Custom neural network

N1 speed Roll

Aircraft gas turbine

N2 speed Vertical acceleration health monitoring

+ ×
휇
−
Total air temperature Product Sqrt
Pitch 2 Add Mean Aircraft gas turbine
performance deterioration value
100
EGT
Angle of attack Ground/air Alarm limit

Figure 10: Aircraft gas turbine engine health monitoring models.

International Journal of Aerospace Engineering 11

CBM: Condition-based maintenance

EGT: Exhaust gas temperature
EHM: Engine health monitoring
FOD: Foreign object damage
MAPE: Mean absolute percentage error
MRA: Multiple regression analysis
MSE: Mean square error
R2 : R squared
RMSE: Root mean square error
TAT: Total air temperature.

Conflicts of Interest
Figure 11: Results.
The authors declare that they have no conflicts of interest.

Acknowledgments
This work was supported by Research Fund of the Erciyes
University, Project no. FDK-2016-6803.

References
Figure 12: Graphical representation of an unusual EGT increase.
6. Conclusions
In this study, we aimed to develop a predictive model of the
EGT parameter using statistical and artificial intelligence
methods from the actual data recorded from the takeoff to
the landing of an aircraft. Statistical multiple regression anal-
ysis and artificial neural networks were used in the prediction
model and the method that gave the lowest error perfor-
mance was tried to be determined. The R2, MAPE, and RMSE
values between the measured values and the predicted values
for the training, validation, and test data of the EGT param-
eter are presented in Tables 11 and 13. When we look at the
test-stage determination coefficient and error rates between
the EGT values from the aircraft’s sensor and the predicted
EGT data, it is seen that the best end result is obtained by
ANN. With the interface we developed in MATLAB Simu-
link, a model has been developed to instantly monitor the
flight status of the aircraft.
In this study, it has been shown that any deterioration of
performance can be easily detected by predicting the EGT
parameter at the takeoff, cruise, and landing phases of the air-
craft. Also, even if there is no fault in the aircraft’s engine, a
fault in the EGT sensor can jeopardize flight safety. It is
thought that the pilot can fly safely with the predicted EGT
value in the developed interface.
[10] T. Brotherton, G. Jahns, J. Jacobs, and D. Wroblewski, “Prog-
Abbreviations
ANN: Artificial neural network
AOA: Angle of attack
[1] A. J. Volponi, “Gas turbine engine health management: past,
present, and future trends,” Journal of Engineering for Gas
Turbines and Power, vol. 136, no. 5, article 051201, 2014.
[2] S. H. Kim, M. A. Cohen, and S. Netessine, “Performance
contracting in after-sales service supply chains,” Management
Science, vol. 53, no. 12, pp. 1843–1858, 2007.
[3] M. J. Dennis and A. Kambil, “Service management: building
profits after the sale,” Supply Chain Management Review,
vol. 7, no. 3, pp. 42–48, 2003.
[4] B. Yu, D. Liu, and T. Zhang, “Fault diagnosis for micro-gas
turbine engine sensors via wavelet entropy,” Sensors, vol. 11,
no. 12, pp. 9928–9941, 2011.
[5] S. Garg, K. Schadow, W. Horn, H. Pfoertner, and I. Stiharu,
“Sensor and actuator needs for more intelligent gas turbine
engines,” ASME Turbo Expo 2010: Power for Land, Sea, and
Air, 2010, pp. 155–167, Glasgow, UK, June 14–18, 2010.
[6] H. Xu, D. Jiang, and L. Liang, “Application of fuzzy-rough set
theory and improved SMO algorithm in aircraft engine vibra-
tion fault diagnosis,” in Proceedings of the IEEE 2012 Prognos-
tics and System Health Management Conference (PHM-2012
Beijing), pp. 1–6, Beijing, China, 2012.
[7] S. M. Lee, T. S. Roh, and D. W. Choi, “Defect diagnostics of
SUAV gas turbine engine using hybrid SVM-artificial neural
network method,” Journal of Mechanical Science and Technol-
ogy, vol. 23, no. 2, pp. 559–568, 2009.
[8] M. A. Zaidan, A. R. Mills, R. F. Harrison, and P. J. Fleming,
“Gas turbine engine prognostics using Bayesian hierarchical
models: a variational approach,” Mechanical Systems and Sig-
nal Processing, vol. 70-71, pp. 120–140, 2016.
[9] A. K. Jardine, D. Lin, and D. Banjevic, “A review on machinery
diagnostics and prognostics implementing condition-based
maintenance,” Mechanical Systems and Signal Processing,
vol. 20, no. 7, pp. 1483–1510, 2006.
nosis of faults in gas turbine engines. In aerospace conference
proceedings, 2000,” in 2000 IEEE Aerospace Conference. Pro-
ceedings (Cat. No.00TH8484), vol. 6, pp. 163–171, Big Sky,
MT, USA, 2000.
12 International Journal of Aerospace Engineering

[11] X. Wang, Aircraft Fuel System Prognostics and Health Manage- [27] A. Patterson-Hine, G. Biswas, G. Aaseng, S. Narasimhan, and
ment [M.S. Thesis], Cranfield University, Bedfordshire, UK, K. Pattipati, “A review of diagnostic techniques for ISHM
2012. applications,” in 1st integrated systems health engineering and
[12] X. Zhang, L. Tang, and J. Decastro, “Robust fault diagnosis of management forum, 2005, Napa, CA, USA, 2005.
aircraft engines: a nonlinear adaptive estimation-based [28] Federal Aviation Administration, “Airplane turbofan engine
approach,” IEEE Transactions on Control Systems Technology, operation and malfunctions basic familiarization for flight
vol. 21, no. 3, pp. 861–868, 2013. crews,” 2016, http://www.faa.gov/aircraft/air_cert/design_
[13] B. Pourbabaee, N. Meskin, and K. Khorasani, “Sensor fault approvals/engine_prop/media/engine_malf_famil.pdf.
detection, isolation, and identification using multiple-model- [29] A. Beldjazia and D. Alatou, “Precipitation variability on the
based hybrid Kalman filter for gas turbine engines,” IEEE massif forest of Mahouna (North Eastern-Algeria) from 1986
Transactions on Control Systems Technology, vol. 24, no. 4, to 2010,” International Journal of Management Sciences and
pp. 1184–1200, 2016. Business Research, vol. 5, no. 3, 2016.
[14] X. F. Zhao, Y. B. Liu, and X. He, “Fault diagnosis of gas turbine
based on fuzzy matrix and the principle of maximum member-
ship degree,” Energy Procedia, vol. 16, pp. 1448–1454, 2012.
[15] N. H. Afgan, M. G. Carvalho, P. A. Pilavachi, A. Tourlidakis,
G. G. Olkhonski, and N. Martins, “An expert system concept
for diagnosis and monitoring of gas turbine combustion
chambers,” Applied Thermal Engineering, vol. 26, no. 7,
pp. 766–771, 2006.
[16] F. Lu, J. Huang, and Y. Lv, “Gas path health monitoring for a
turbofan engine based on a nonlinear filtering approach,”
Energies, vol. 6, no. 1, pp. 492–513, 2013.
[17] M. Amozegar and K. Khorasani, “An ensemble of dynamic
neural network identifiers for fault detection and isolation of
gas turbine engines,” Neural Networks, vol. 76, pp. 106–121,
2016.
[18] D. Zhou, H. Zhang, and S. Weng, “A new gas path fault
diagnostic method of gas turbine based on support vector
machine,” Journal of Engineering for Gas Turbines and Power,
vol. 137, no. 10, article 102605, 2015.
[19] R. Chacartegui, D. Sánchez, A. Munoz, and T. Sánchez, “Real
time simulation of medium size gas turbines,” Energy Conver-
sion and Management, vol. 52, no. 1, pp. 713–724, 2011.
[20] S. Ackert, Elements of Turbofan Shop Maintenance Costs,
p. 1–42, 2011.
[21] M. T. Yildirim and B. Kurt, “Engine health monitoring in an
aircraft by using Levenberg-Marquardt feedforward neural
network and radial basis function network,” in 2016 Interna-
tional Symposium on INnovations in Intelligent SysTems and
Applications (INISTA), pp. 1–5, Sinaia, Romania, 2016.
[22] H. A. Es, F. Y. Kalender, and C. Hamzacebi, “Forecasting the
net energy demand of Turkey by artificial neural networks,”
Journal of the Faculty of Engineering and Architecture of
Gazi University, vol. 29, no. 3, pp. 495–504, 2014.
[23] B. Rosner, Fundamentals of Biostatistics, Brooks/Cole, Bos-
ton, MA, USA, 7th edition, 2010.
[24] C. D. Lewis, Industrial and Business Forecasting Methods,
Butterworths Publishing, London, 1982.
[25] Ş. Yurtcu and A. Özocak, “Prediction of compression index of
fine-grained soils using statistical and artificial intelligence
methods,” Journal of the Faculty of Engineering and Architec-
ture of Gazi University, vol. 31, no. 3, pp. 597–608, 2016.
[26] A. Babbar, V. L. Syrmos, E. M. Ortiz, and M. M. Arita,
“Advanced diagnostics and prognostics for engine health
monitoring,” in 2009 IEEE Aerospace conference, pp. 1–10,
Big Sky, MT, USA, 2009.
 International Journal of

Rotating Advances in
Machinery Multimedia

The Scientific
Engineering
Journal of
Journal of

Hindawi
World Journal
Hindawi Publishing Corporation Hindawi
Sensors
Hindawi Hindawi
www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi
Engineering
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Volume 2018
Navigation and
Observation
Hindawi
www.hindawi.com Volume 2018
in Engineering
Hindawi
www.hindawi.com Volume 2018
Engineering
Hindawi
www.hindawi.com Volume 2018
Hindawi
www.hindawi.com www.hindawi.com Volume 2018

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Hindawi Hindawi Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018